Fed-RAM: Federated Learning-Based Robust and Adaptive Method for Anomaly Detection in Industrial IoT Systems

Traditional machine learning (ML) approaches for anomaly detection rely on centralized data and struggle to manage the complexity, privacy, and scale of data generated by the industrial Internet of Things (IIoT). Recently, Federated learning (FL) offers a viable solution by enabling diverse IIoT devices and smart machines to collaboratively train high-quality ML models without transmitting sensitive data to ensure privacy and data security. However, the widespread adoption of FL in heterogeneous IIoT systems is primarily caused by model training inefficiencies, delays, and communication overhead, specifically for time-sensitive IIoT applications. With this focus, in this letter, an FL-based robust and adaptive method for anomaly detection, Fed-RAM, is proposed for edge-assisted IIoT. Specifically, we design adaptive mechanisms aiming to improve anomaly detection through an adaptive anomaly score technique based on the reconstruction error that enables local models to dynamically adjust the anomaly threshold according to the unique data distribution of individual edge devices. By incorporating adaptive mechanisms at each edge device, Fed-RAM ensures that only significant and critical updates are communicated to the server for aggregation. The experimental results obtained on two widely known datasets, i.e., MNIST and CIFAR-10, demonstrate that Fed-RAM reduces communication costs by approximately 50<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> and improves anomaly detection accuracy by 14<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> compared to baseline models.